Scholarlink

<p>The calibration of microparameters in particle flow code (PFC) remains a major bottleneck due to their non-physical nature and the inefficiency of conventional trial-and-error methods. To address these limitations, this study proposes a novel, efficient, and interpretable calibration framework by integrating a Particle Swarm Optimization-optimized Backpropagation Neural Network (PSO-BP) surrogate model for inverse parameter identification in PFC. By leveraging 300 automated uniaxial compression simulations, a robust nonlinear mapping from micro- to macro-parameters is established, achieving exceptional predictive accuracy (<i>R</i><sup>2</sup> &gt; 0.99). The optimized surrogate enables inversion within ~ 5&#xa0;min, with errors as low as 1.6% (uniaxial strength) and 1.8% (elastic modulus) against experimental data. Remarkably, the model exhibits outstanding generalizability across six rock types, yielding an average accuracy of 99%. For the first time in PFC calibration, SHAP interpretability analysis is employed to unveil the physical significance of microparameters, identifying <i>emod</i> and <i>krat</i> as dominant for elastic modulus, and <i>pb_coh</i> and <i>pb_ten</i> as primary controls for uniaxial strength.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A High-Efficiency and Interpretable Framework for Microparameter Calibration via PSO-Optimized Machine Learning and SHAP Analysis

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Jing Zhou,
  • Hang Lin,
  • Yifan Chen

摘要

The calibration of microparameters in particle flow code (PFC) remains a major bottleneck due to their non-physical nature and the inefficiency of conventional trial-and-error methods. To address these limitations, this study proposes a novel, efficient, and interpretable calibration framework by integrating a Particle Swarm Optimization-optimized Backpropagation Neural Network (PSO-BP) surrogate model for inverse parameter identification in PFC. By leveraging 300 automated uniaxial compression simulations, a robust nonlinear mapping from micro- to macro-parameters is established, achieving exceptional predictive accuracy (R2 > 0.99). The optimized surrogate enables inversion within ~ 5 min, with errors as low as 1.6% (uniaxial strength) and 1.8% (elastic modulus) against experimental data. Remarkably, the model exhibits outstanding generalizability across six rock types, yielding an average accuracy of 99%. For the first time in PFC calibration, SHAP interpretability analysis is employed to unveil the physical significance of microparameters, identifying emod and krat as dominant for elastic modulus, and pb_coh and pb_ten as primary controls for uniaxial strength.